Model Risk Cultures
Dr Andreas Tsanakas Cass Business School, City University London
Annual Meeting of the Swiss Association of Actuaries
Davos, 5 September 2014
Acknowledgments
● M Bruce Beck, Warnell School of Forestry and Natural
Resources, University of Georgia
● Michael Thompson, International Institute for Applied
Systems Analysis
● The Model Risk Working Party of the Institute and
Faculty of Actuaries
What is model risk? Office of the Comptroller of the Currency (2011)
[T]he potential for adverse consequences from decisions
based on incorrect or misused model outputs and reports.
Model risk occurs primarily for two reasons:
● The model may have fundamental errors and may
produce inaccurate outputs when viewed against the
design objective and intended business uses. […]
● The model may be used incorrectly or
inappropriately.
Questions
● Are there model risks not associated with model
error?
● How do different ways of using models generate
model risks?
● What is the upside of model risk?
● What sorts of governance responses do different
model uses/risks necessitate?
● And what of Risk Culture?
Types of model risk
● Technical
○ Coding errors, approximations
○ Arising from model uncertainty
● Behavioural
○ Information rejection
○ Loss of accountability
● Systemic
○ Market use of same models (CAT models, ESGs)
○ Endogenous model risks (VaR trading limits)
○ Can also be result of model-free strategies
Model uncertainty (A)
A. Model uncertainty refers to all the possible ways in
which the model used diverges from some true but
unknown model of the process under consideration
● Under definition A., we cannot know the extent of
model uncertainty, but we can sometimes make
informed guesses
Model uncertainty (B)
B. Model uncertainty refers to all the potential ways in
which the model’s inputs, outputs and structure may
change under plausible changes in assumptions
● Under definition B., the extent of model uncertainty is
revealed by sensitivity analysis
○ Materiality is application specific
○ Plausibility is not only a technical matter
○ See Beck (2014) for in-depth discussion
Sensitivity of annuity value to model choice (70 year old male, discount at 3%; Richards et al, 2013)
0
0.5
1
1.5
2
2.5
11 12 13 14
Pro
bab
ility
Den
sity
Annuity Value
LC
DDE
LC(S)
CBD Gompertz
CBD Pspline
APC
2DAP
Perception of models
Confidence / low concern for model uncertainty
Diffidence / high concern for model uncertainty
High legitimacy
of modelling
Low legitimacy
of modelling
Perception of models
Confident model users
Known knowns
Confidence / low concern for model uncertainty
Diffidence / high concern for model uncertainty
High legitimacy
of modelling
Low legitimacy
of modelling
Perception of models
Confident model users
Known knowns
Conscientious modellers
Known unknowns
Confidence / low concern for model uncertainty
Diffidence / high concern for model uncertainty
High legitimacy
of modelling
Low legitimacy
of modelling
Perception of models
Confident model users
Known knowns
Uncertainty avoiders
Unknown unknowns
Conscientious modellers
Known unknowns
Confidence / low concern for model uncertainty
Diffidence / high concern for model uncertainty
High legitimacy
of modelling
Low legitimacy
of modelling
Perception of models
Expert decision makers
Unknown knowns
Confident model users
Known knowns
Uncertainty avoiders
Unknown unknowns
Conscientious modellers
Known unknowns
Confidence / low concern for model uncertainty
Diffidence / high concern for model uncertainty
High legitimacy
of modelling
Low legitimacy
of modelling
Confident model users: optimality
● Model drives decisions
● Focus on using model to
identify and exploit
opportunities
● Information discordant with
model is ignored
● Risk: world behaves very
differently to what model
predicts
Expert decision makers
Confident model users
Uncertainty avoiders
Conscientious modellers
Conscientious modellers: fitness for purpose
● Model may inform some
decisions but not others
● Focus on stating uncertainties /
delimiting fitness of purpose
● While uncertainty is high,
overall paradigm is appropriate
● Risk: constraints on model
use / suboptimal strategy
● Risk: paradigm is wrong
Expert decision makers
Confident model users
Uncertainty avoiders
Conscientious modellers
Uncertainty avoiders: robustness
● Models are never good enough
to drive decisions
● Focus on adverse scenarios,
structural changes,
interconnectedness of risks
● Conservative strategies that are
robust to model uncertainty
● Risk: highly suboptimal
decisions
Expert decision makers
Confident model users
Uncertainty avoiders
Conscientious modellers
Expert decision makers: intuition
● Decisions driven by
management expertise alone
[But model may be manipulated
to produce convenient answers]
● Information discordant with
intuition is ignored
● Risk: missing dangers and
opportunities that model
could identify
● [Risk: loss of accountability]
Expert decision makers
Confident model users
Uncertainty avoiders
Conscientious modellers
On individuals
● The classification reflects ways of thinking about
models, not individual psychological profiles
● People may hold different views depending e.g. on
context and professional affiliation
○ Or indeed entertain conflicting views
● Each perspective challenges the other three
○ Twelve challenges in total
On individuals
● The classification reflects ways of thinking about
models, not individual psychological profiles
● People may hold different views depending e.g. on
context and professional affiliation
○ Or indeed entertain conflicting views
● Each perspective challenges the other three
○ Twelve challenges in total
● People may be responsive to those who hold different
views or instead...
What “we” may say about “them”
Expert decision makers
Confident model users
“Naïve”
Uncertainty avoiders
Conscientious modellers
What “we” may say about “them”
Expert decision makers
“Stone-age cynics”
Confident model users
Uncertainty avoiders
Conscientious modellers
What “we” may say about “them”
Expert decision makers
Confident model users
Uncertainty avoiders
“Unhelpful scaremongers”
Conscientious modellers
What “they” may say about “us”
Expert decision makers
Confident model users
Uncertainty avoiders
Conscientious
modellers
“Obstructive
worriers”
What “they” may say about “us”
Expert decision makers
Confident model users
Uncertainty avoiders
Conscientious
modellers
“Irrelevant
navel-gazers”
What “they” may say about “us”
Expert decision makers
Confident model users
Uncertainty avoiders
Conscientious
modellers
“Missing the
forest for the
tree”
Model specification and delivery (“us” vs. Confident model users)
● Technical modellers now recognize how model
complexity (“granularity”) increases sensitivity
○ E.g. in hierarchical dependence modelling
● Management information, reporting requirements,
model embedding, require more detail
● Statistical inference, sensitivity analysis, may point out
the need for less detail
Paradigm challenge (“us” vs. Uncertainty avoiders)
● Economic theory is flawed (rational agents,
frictionless markets, linear valuation functionals...)
● Endogeneity of financial risk
● Claims generating processes undergo unobservable
structural changes
● But how can we quantify uncertainty without employing
a paradigm?
● How can we think outside the box, without a box?
Input/output feedback (“us” vs. Expert decision makers)
● We know that portfolio risk can greatly vary with small
changes in hard-to-validate dependence assumptions
● Statistically plausible risk metrics (reserve estimates,
portfolio VaRs...), when linked to decisions, can be
commercially unviable
● Model uncertainty allows choice of dependence
parameters that give reasonable outputs
● Reasonableness is a social construct
Input/output feedback (“us” vs. Expert decision makers)
“Both groups have difficulty recognizing the ways in which
the process subtly interweaves truth seeking and
advantage seeking, leaving each somewhat compromised
by the other, even as each somewhat serves the other.”
(March, 1994)
Basket of 5 defaultable securities p=5%, Beta-binomial dependence
Number of
Defaults
Probability
(corr=0)
Multiplier
(corr=1%)
Multiplier
(corr=5%)
Multiplier
(corr=10%)
5 0.0000313% 4.44 84.68 508.53
4 0.00297% 2.45 16.25 51.42
3 0.113% 1.54 4.13 7.49
2 2.14% 1.12 1.48 1.70
1 20.4% 0.96 0.84 0.71
0 77.4% 1.01 1.02 1.04
Hegemony eats itself
“Why didn't rating agencies build in some cushion for this
sensitivity to a house-price-depreciation scenario? Because if
they had, they would have never rated a single mortgage-
backed CDO.” (Salmon, 2009)
“Where is the liquidity crisis supposed to come from?”
somebody asked in the meeting. No one could give a good
answer. [...] In their eyes, we were not earning money for the
bank. Worse, we had the power to say no and therefore prevent
business from being done. Traders saw us as obstructive and
a hindrance to their ability to earn higher bonuses.”
(Confessions of a risk manager, Economist, 2008)
Echoes of Cultural Theory Thompson et al (1990); Ingram et al (2012)
Echoes of Cultural Theory Thompson et al (1990); Ingram et al (2012)
1. In the presence of deep uncertainties, risk is socially
constructed. What we consider to be risky depends on
moral and political considerations.
Echoes of Cultural Theory Thompson et al (1990); Ingram et al (2012)
1. In the presence of deep uncertainties, risk is socially
constructed. What we consider to be risky depends on
moral and political considerations.
2. There are four fundamental ways of perceiving risk
(rationalities / Risk Cultures), linked to different ways of
organising social and economic relations (ways of life).
Echoes of Cultural Theory Thompson et al (1990); Ingram et al (2012)
1. In the presence of deep uncertainties, risk is socially
constructed. What we consider to be risky depends on
moral and political considerations.
2. There are four fundamental ways of perceiving risk
(rationalities / Risk Cultures), linked to different ways of
organising social and economic relations (ways of life).
3. Each of these Risk Cultures represents a distinct form of
human wisdom. But they are also mutually incompatible –
consensus is not possible.
Echoes of Cultural Theory Thompson et al (1990); Ingram et al (2012)
1. In the presence of deep uncertainties, risk is socially
constructed. What we consider to be risky depends on
moral and political considerations.
2. There are four fundamental ways of perceiving risk
(rationalities / Risk Cultures), linked to different ways of
organising social and economic relations (ways of life).
3. Each of these Risk Cultures represents a distinct form of
human wisdom. But they are also mutually incompatible –
consensus is not possible.
4. An institution committed to only one of these Risk Cultures
is not viable. Each way of life needs the presence of all
others to be viable.
Myths of nature
“Irrelevant
navel-gazers”
Nature benign
Known knowns
Nature capricious
Unknown knowns
Nature perverse/tolerant
Known unknowns
Nature ephemeral
Unknown unknowns
Clumsy solutions
● Lack of consensus means that viable solutions will
have to be clumsy rather than elegant
● Different rationalities need to be represented, have
access to decision making and be responsive to
each other
Clumsy solutions
● Lack of consensus means that viable solutions will
have to be clumsy rather than elegant
● Different rationalities need to be represented, have
access to decision making and be responsive to
each other
● The people of Davos have managed such
pluralism for many years (Thompson, 2002)
○ Private property
○ Wald- and Alpgenossenschaften
○ Avalanche control
○ Casual labour
Model governance
● Current best-practice
○ Model inventory and documentation standards
○ Improved articulation of model limitations
○ Requirements and restrictions on model use
○ Independent model validation process
Model governance
● Current best-practice
○ Model inventory and documentation standards
○ Improved articulation of model limitations
○ Requirements and restrictions on model use
○ Independent model validation process
● “Culturally aware governance”
○ Multi-way challenge
○ Accessibility and responsiveness
Regulatory guidance: 1-way challenge
Expert decision makers
Confident model users
Uncertainty avoiders
Office of the Comptroller of the
Currency
Regulatory guidance: 1-way challenge
Expert decision makers
Confident model users
Conscientious modellers
Haldane & Madouros:
The Dog and the Frisbee
Challenges to “conscientious modelling”
● Conditions for embedding in decisions and attracting
investment in the model
○ Model user friendly, addresses business issues,
released on time
○ Test output against expert opinion and
commercial implications
○ Extended peer review
Challenges to “conscientious modelling”
● Conditions for embedding in decisions and attracting
investment in the model
○ Model user friendly, addresses business issues,
released on time
○ Test output against expert opinion and
commercial implications
○ Extended peer review
● But we also need strong challenges to model-bashing
and manipulation
The quest for openness
● Should we (?) be able to openly say this:
“Assumptions are consistent with empirical evidence
and best modelling practice. Model uncertainty
remains high. The precise model calibration is such
that standard outputs are also consistent with
senior management’s perspective of a commercially
reasonable capital requirement.”
The quest for openness
● Should we (?) be able to openly say this:
“Assumptions are consistent with empirical evidence
and best modelling practice. Model uncertainty
remains high. The precise model calibration is such
that standard outputs are also consistent with
senior management’s perspective of a commercially
reasonable capital requirement.”
● If we acknowledge the legitimacy of such concerns,
will this improve scientific integrity?
○ To ask hard questions, we need to be unafraid of
the answers
Yes but...
● Why would Expert Decision Makers choose to
manipulate models
○ Why use a model at all?
● In the face of deep uncertainties, what exactly do
Conscientious Modellers have to offer?
○ Apart from: “sorry can’t answer this question”...
Yes but...
● Why would Expert Decision Makers choose to
manipulate models
○ Why use a model at all?
● In the face of deep uncertainties, what exactly do
Conscientious Modellers have to offer?
○ Apart from: “sorry can’t answer this question”...
● If we had better data / models / software etc, could
we all agree what to do?
Decision principles
● To make a decision we need
○ Technical (probability) assessment
○ Decision principle
● Disagreement can be in respect of either element
● Sensitivity of decision to model changes reveals
materiality of model uncertainty
● Rigid application of some sanctioned decision principle
is a form of hegemony
Fields of contestation (how we often think)
Science
contested
Principle
contested
Simple problems No No
Model risk Yes No
Ecstasy v horse-riding No Yes
Climate change Yes Yes
Fields of contestation (how things are)
Science
contested
Principle
contested
Simple problems No No
Model risk Yes No Yes
Ecstasy v horse-riding No Yes Yes
Climate change Yes Yes
Rebellious rationalities
● Conscientious modellers
○ Re-define purpose of modelling!
○ (Protect integrity and need for expertise)
● Uncertainty avoiders
○ Change decision principle!
○ (Protect enterprise or market)
● Expert decision makers
○ Rig the model to get the answers I want!
○ (Survive by manipulation)
Conclusions
● Governance should encourage multi-way challenges
to model (non-)use as legitimate
○ Are all perspectives heard?
○ Do all perspectives respond to others?
● Not easy to distinguish contestation of models from
contestation of decision principles
● How does model governance improve accountability?
● Acknowledging the political nature of risk
○ Less politics worse science
References
1. Beck (2014), “Handling uncertainty in environmental models at the
science-policy-society interfaces,” in Boumans, Hon, and Petersen,
(eds), Error and Uncertainty in Scientific Practice, Pickering &
Chatto, pp 97-135.
2. Haldane and Madouros (2012) “The dog and the frisbee”, Federal
Reserve Bank of Kansas City’s 36th economic policy symposium.
3. Ingram, Tayler, and Thompson (2012), “Surprise, surprise: from
neoclassical economics to e-life”, ASTIN Bulletin 42 (2), 389-411.
4. March (1994), A Primer on Decision Making: How Decisions
Happen, Simon & Schuster.
References
5. Office of the Comptroller of the Currency (2012), “Supervisory
guidance on model risk management”, Report OCC 2011-22,
Board of Governors of the Federal Reserve System
6. Richards, Currie, and Ritchie (2013) “A Value-at-Risk framework
for longevity trend risk” British Actuarial Journal 19(01), 116-139.
7. Thompson (2002), “Man and nature as a single but complex
system,” in Munn (ed), Encyclopedia of Global Environmental
Change, Vol 5, Wiley, pp 384-393.
8. Thompson, Ellis, and Wildavsky (1990), Cultural Theory, Westview
Press.
Top Related